Rapid recent progress in data-driven forecasting using machine learning (ML) has shown the potential to revolutionize weather and climate prediction.
Trained on large-scale, high-resolution datasets, ML-based models can now rival or even surpass the performance of traditional physics-based numerical weather prediction models, particularly global forecasting. However, high-resolution, regional data-driven models remain an area of ongoing research. This research proposal focuses on developing a skillful high-resolution, data-driven forecasting model for the Alpine region using innovative ML methods that provide the most accurate forecast.
The system will be trained on global reanalysis data, a unique 1km high-resolution regional dataset over the Alpine domain with 20 years of data. Through a multi-stage transfer learning protocol, we intend to produce models that outperform conventional NWP in forecast accuracy, temporal resolution, and computational efficiency. To accomplish these objectives, we request computational and storage resources to support the multi-stage training, fine-tuning, and validation of these models.
The main project outcomes will be peer-reviewed scientific publications, model datasets, open model weights, and open-source code repositories that enable replication and further exploration by the broader research community.
In summary, this project addresses critical gaps in regional data-driven weather modeling by exploring ML techniques capable of high-resolution forecasting for complex Alpine terrain, with implications for advancing both scientific understanding and practical applications in similar regions worldwide.
Carlos Osuna, MeteoSwiss, Switzerland